News
Kubernetes with machine learning as a crystal ball
Kubernetes can automatically scale applications, but in principle it only responds to existing demand. The Dutch Education Executive Agency (DUO) is exploring how KEDA and machine learning can transform this into a predictive model. Instead of scaling infrastructure only when demand increases, the platform anticipates it. During Edgecase 2026, Rick Brouwer and Harke Wijnsma, both Practice Leads OPS at DUO, demonstrated the ins and outs of predictive Kubernetes Event-Driven Autoscaling.
According to Brouwer and Wijnsma, KEDA is typically based on historical data and operates largely in the present. However, it can also be a powerful tool for predicting the future. Through two demonstrations, they showed how this works in practice.
Being passionate technologists, they also added a playful twist. Using Minikube and KEDA, they created a setup that generated Pokémon on a screen. The louder the audience cheered, the more Pokémon appeared. In Hilversum, that proved highly effective. “I’ve never counted that many before!” Brouwer remarked enthusiastically. It was a literal example of an event-driven approach, jokingly dubbed CEDA: Community Event-Driven Autoscaling.
The conductor
Before introducing the predictive model, the speakers demonstrated KEDA, with the “K” standing for Kubernetes. The key ingredient, as Wijnsma explained, is machine learning.
How valuable would it be if Kubernetes could already know that system load will increase in ten or twenty minutes? At DUO, historical data is used to identify demand patterns and generate forecasts of expected workload. For this purpose, Elasticsearch is used to analyze trends and predict future demand.
KEDA can then use these forecasts to proactively provision the required capacity. Instead of adding pods after demand has already increased, they are started in advance. The Elastic Forecast Scaler orchestrates these scaling decisions.
This final step is what makes KEDA particularly interesting, Brouwer and Wijnsma argued. The prediction does not remain isolated in a monitoring dashboard but becomes an active part of the operational infrastructure itself.
Edgecase 2026: Gotta read 'em all
All the speakers, sessions, and insights from Edgecase 2026 in one place.
Safety net
With this approach, the role of operators and platform teams begins to change. Rather than manually defining scaling rules for every application and expected peak, teams can rely on a system that anticipates demand automatically.
That does not mean reactive autoscaling disappears. Even the most accurate forecasts cannot eliminate unexpected spikes or sudden drops in demand. A safety net is therefore still required. Traditional autoscaling remains an essential fallback mechanism alongside predictive scaling.
